Ananke.Orchestration.Knowledge 0.6.0

Ananke.Orchestration.Knowledge

Knowledge pipeline for Ananke — vector knowledge stores, document processing, chunking, embedding abstractions, knowledge catalog with LLM-enriched metadata, and cross-document linking.

Install

dotnet add package Ananke.Orchestration.Knowledge

Note: If you use Ananke.Orchestration, this package is already included as a transitive dependency.

Quick start

Semantic search

using Ananke.Orchestration.Knowledge;

// InMemoryKnowledgeStore for dev/test; QdrantKnowledgeStore for production
var store = new InMemoryKnowledgeStore(embeddingModel);

await store.UpsertAsync([
    new KnowledgeDocument { Id = "doc-1", Text = "Ananke is a .NET agent orchestration library." },
    new KnowledgeDocument { Id = "doc-2", Text = "Workflows are built with a fluent graph-as-code API." },
]);

var results = await store.SearchAsync("how do I build a workflow?");
// results[0].Text => "Workflows are built with a fluent graph-as-code API."

Document ingestion pipeline

using Ananke.Orchestration.Knowledge;
using Ananke.Orchestration.Knowledge.Documents;

var processor = new DocumentProcessor(
    new HttpClient(),
    [new PdfExtractor(), new MarkdownExtractor()],  // from Ananke.Documents
    new SlidingWindowChunker(),
    knowledgeStore);

await using var stream = File.OpenRead("architecture.pdf");
var result = await processor.ProcessAsync(stream, "application/pdf", sourceId: "arch-doc");
// result.ChunkCount => 42

Knowledge catalog (two-phase discovery)

using Ananke.Orchestration.Knowledge.Catalog;

var catalog = new InMemoryKnowledgeCatalog(embeddingModel);

// Wrap any store to auto-update catalog on upsert/delete
var catalogStore = new CatalogAwareKnowledgeStore(innerStore, catalog, enricher);

// Phase 1: discover relevant sources
var sources = await catalog.BrowseAsync(new CatalogBrowseOptions { Query = "deployment" });

// Phase 2: deep-search within matched sources
var chunks = await catalogStore.SearchAsync("kubernetes deployment", new SearchOptions
{
    Filter = new KnowledgeFilter { ["source_id"] = sources[0].Entry.SourceId }
});

Cross-document linking

using Ananke.Orchestration.Knowledge.Linking;

// Opt-in via DI — decorates IKnowledgeStore with graph-expanded search
services.AddKnowledgeLinking(options =>
{
    options.AutoLinkOnIngest = true;       // LLM-based post-ingestion linking
    options.SimilarityThreshold = 0.7f;    // link documents above this score
});

Key types

The pipeline

1. IDocumentExtractor.ExtractAsync(stream) → structured text + sections
2. IDocumentChunker.Chunk(text)            → DocumentChunk[]
3. IEmbeddingModel.EmbedBatchAsync(chunks) → float[][] vectors
4. IKnowledgeStore.UpsertAsync(documents)  → indexed in vector store
5. IKnowledgeStore.SearchAsync(query)      → ranked KnowledgeChunk[]

Optional enrichment layers compose via decoration:

IKnowledgeStore
  └─ CatalogAwareKnowledgeStore  (auto-catalogs on upsert)
      └─ LinkedKnowledgeStore    (graph-expanded search)

Implementations

Dependencies

• Ananke.Abstractions — for IAgentModel, IEmbeddingModel, AgentMessage

This package has no dependency on Ananke.Orchestration — it is independent of the workflow engine.

Related packages

License

Apache 2.0

• Markdig ≥ 0.40.0 (transitive)

Related packages

Documentation

Full docs, demos, and architecture: github.com/sevensamurai/Ananke

License

Apache 2.0